chore(v1.10): price-table + model-alias refresh

[jameskhair-code]

Item 2 of the v1.10 preflight (docs/planning/v1.10-charter.md). Prices verified against the live Anthropic pricing docs (platform.claude.com/docs/en/about-claude/pricing) on 2026-06-10 — not from memory.

What changed

usage.py:_PRICES — two real errors fixed:

• The single claude-opus-4 prefix entry carried $15/$75 (Opus 4.0/4.1-era rates). Opus 4.5 through 4.8 actually bill at $5/$25 (cache read $0.50, 5m cache write $6.25). The table now has per-version entries for claude-opus-4-5/-6/-7/-8 at the new rates; longest-prefix matching keeps the bare claude-opus-4 entry as the $15/$75 fallback for 4.0/4.1 dated IDs (e.g. claude-opus-4-20250514). Net effect: anything billed at latest was being over-estimated 3×.
• claude-haiku-4 carried $0.80/$4.00 — those are Haiku 3.5 rates. Haiku 4.5 is $1/$5 (read $0.10, write $1.25).
• Sonnet 4.x unchanged at $3/$15. The 3.x legacy entries stay (correct as written) so old usage-log replays keep pricing.

models.py — latest rolls claude-opus-4-7 → claude-opus-4-8 (the current Opus; same API surface as 4.7). _DEPRECATED_MODELS seeded from Anthropic's deprecation schedule: claude-opus-4-1 (retires 2026-08-05), claude-opus-4-20250514 / claude-sonnet-4-20250514 (retire 2026-06-15), claude-3-haiku-20240307 (retired). fast/legacy unchanged; no docs/config changes needed — they reference aliases only, and the real config.json pins claude-sonnet-4-6 explicitly so nothing changes behavior today.

The campaign model decision (James picks)

Projected cost for one full sweep of the four measured AI steps over ~5,000 books, from usage.jsonl history (228 calls, all Sonnet 4.6) — per-call averages divided by each step's default batch size, priced at verified rates. The Opus column includes a +30% token adjustment for the 4.7+ tokenizer (docs say "up to 35% more tokens").

Step	per-book tokens (in/out/cache-write/cache-read)	Opus 4.8	Sonnet 4.6	Haiku 4.5
clean-titles	48 / 67 / 119 / 723	$19.64	$9.07	$3.02
lcc-enrich	107 / 86 / 620 / 169	$43.20	$19.94	$6.65
comments-enrich	159 / 503 / 1,205 / 289	$136.75	$63.11	$21.04
tags-enrich	64 / 32 / 54 / 108	$9.87	$4.56	$1.52
Full-library total		≈ $209	≈ $97	≈ $32

Per-book, all four steps: Opus $0.042 · Sonnet $0.019 · Haiku $0.006.

Caveats, stated honestly:

• These skew high. History is dominated by small smoke batches, so fixed per-call overhead and cold cache writes are spread over fewer books than a real 250–500-book wave would amortize. Campaign-scale runs should land below these figures.
• Not covered: the enrich-identifiers AI path (never run — no usage history; applies only to a subset of books) and the tags-cleanup pass (batches of 150, few calls — small). Both add to the total but not materially against the model deltas.
• All history is Sonnet 4.6; cross-model figures assume the same token shape, which is the simplest honest projection available.

My read: comments-enrich is ~2/3 of total cost at any model, and the Opus-vs-Sonnet delta on the whole campaign is ~$110. The decision is recorded when you pick — including per-step mixes (e.g. Sonnet everywhere, Opus for comments only ≈ $170).

Verification

• python -m pytest -q — 639 passed (3 price-pinned tests updated to the verified rates, including a new assertion that legacy Opus 4.0 IDs still resolve to $15).
• Real-library smoke: clean-titles --limit 3 --dry-run shows a dollar estimate (≈ $0.04, cold-cache write dominated) with the refreshed table. A smoke on the chosen campaign model happens after your pick.

🤖 Generated with Claude Code